Anthropic Accuses DeepSeek, Moonshot, and MiniMax of 'Distillation' Campaigns Against Claude AI Model
Introduction
The race for dominance in artificial intelligence between the United States and China has taken a new turn, with the focus shifting from training the most powerful models to the rules of the game. Anthropic has accused DeepSeek, Moonshot, and MiniMax of engaging in 'distillation' campaigns against its Claude AI model, extracting capabilities through millions of queries and interactions.
The Accusation
According to Anthropic, the campaigns involved over 16 million queries, interactions, and question-and-answer sessions, channeled through approximately 24,000 fraudulent accounts. This is a clear violation of Anthropic's terms of service and regional access restrictions.
Context and Implications
The announcement comes amidst growing tension surrounding the progress of Chinese AI. DeepSeek's launch of R1, a competitive model developed at a fraction of the cost of US alternatives, had an immediate impact on markets and rekindled the debate in Washington about technological advantage over China.
Distillation: A Double-Edged Technique
Distillation, in simple terms, involves training a less capable model using the responses generated by a more powerful one. While this technique is commonly used in the industry to create smaller, more economical versions of systems, Anthropic argues that it becomes problematic when used to 'acquire powerful capabilities from other laboratories in a fraction of the time and at a fraction of the cost' of independent development.
A Recognizable Pattern
The three laboratories allegedly used fraudulent accounts and proxy services to access Claude on a large scale, attempting to evade detection systems. Anthropic details the infrastructure, including 'hydra clusters' of accounts distributing traffic between their API and third-party cloud platforms, allowing them to replace blocked accounts with new ones.
Three Campaigns, Different Objectives
Although Anthropic presents the campaigns as part of the same dynamic, there are distinct differences. DeepSeek focused on extracting reasoning capabilities and generating safe alternatives to politically sensitive questions. Moonshot aimed at developing agents capable of using tools and manipulating computer environments. MiniMax, with the largest volume of queries, reacted quickly to the launch of a new system, redirecting traffic to extract capabilities from the latest model.
Conclusion
The accusations by Anthropic highlight the evolving landscape of AI development, where the rules of the game are becoming increasingly important. As the race for AI dominance continues, the focus on ethical development and the protection of intellectual property will become more critical.